Virtual power plant vehicle network resource optimization control method and system based on digital twinning

Abstract

The application discloses a virtual power plant vehicle network resource optimization control method and system based on digital twinning, and belongs to the technical field of intelligent power grid resource optimization. The method comprises the following steps: collecting video stream data and battery state data, extracting physical connection stability features, thermal stress features and energy state features, dynamically clustering electric vehicles to form resource clusters, taking the resource clusters as intelligent agents, constructing a digital twinning collaborative optimization model based on a multi-agent deep reinforcement learning algorithm, and generating charging and discharging control instructions. The application introduces computer vision to evaluate the physical connection reliability, combines a gated recurrent unit network to evaluate the battery health state, realizes comprehensive perception of the vehicle state, and solves the problems of scheduling execution failure and accelerated battery life decay caused by incomplete state evaluation in the prior art through the collaborative optimization of dynamic clustering and multi-agent deep reinforcement learning, thereby improving the safety, accuracy and reliability of vehicle network resource optimization control.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent power grid resource optimization, and particularly relates to a virtual power plant vehicle network resource optimization control method and system based on digital twinning. BACKGROUND

[0002] Virtual power plant vehicle network resources refer to electric vehicles that are geographically dispersed and have bidirectional interaction capability with the power grid, which are aggregated into a unified and schedulable collection through advanced communication and control technology. The core component of virtual power plant vehicle network resources is the energy storage system of the electric vehicle battery. Virtual power plant vehicle network resources have adjustable charging load and controllable discharging power. By centrally monitoring and optimizing the scheduling of the aggregated electric vehicle cluster, the virtual power plant vehicle network resources participate in the power market and grid auxiliary services as a whole to provide capacity support, peak and frequency regulation services, and backup capacity for the power system, thereby improving the flexibility and reliability of power grid operation.

[0003] In order to solve the problem of accurate control and safe operation of large-scale electric vehicle clusters participating in grid scheduling, the existing technology uses the battery electrical state parameters obtained through vehicle network communication to establish an aggregation optimization model for unified scheduling processing. However, it cannot quantitatively evaluate the physical connection state of the charging interface, and ignores the differences in the health status of the electric vehicle batteries within the cluster, which leads to the problem of scheduling failure and accelerated battery life decay caused by issuing scheduling instructions to vehicles with unreliable physical connections and performing non-differential charging and discharging control on batteries with different health states. SUMMARY

[0004] The application aims to provide a virtual power plant vehicle network resource optimization control method and system based on digital twinning, which improves the safety, accuracy and reliability of vehicle network resource optimization control.

[0005] To solve the above technical problems, the technical solution adopted by the application is as follows:

[0006] The virtual power plant vehicle network resource optimization control method based on digital twinning comprises the following steps:

[0007] Collect and pre-process the video stream data of the charging port area and the battery state data of the aggregated electric vehicles to obtain pre-processed video stream data and battery state data;

[0008] Input the pre-processed video stream data into a three-dimensional convolutional neural network model to extract physical connection stability features and establish an early warning mechanism to extract thermal stress features and energy state features from the pre-processed battery state data;

[0009] The thermal stress feature and the energy state feature are input into a gated recurrent unit network model, and a health state parameter of the battery is output;

[0010] According to the physical connection stability feature and the health state parameter, a density-based noise application spatial clustering algorithm is adopted to divide the accessed electric vehicles into different resource clusters;

[0011] The aggregated parameters of the resource clusters are taken as inputs of a digital twin collaborative optimization model based on a multi-agent deep reinforcement learning algorithm, distributed collaborative solving is performed, a total power adjustment instruction for each resource cluster is obtained, and power secondary distribution is performed to generate a charging and discharging control instruction;

[0012] The physical connection stability feature, the health state parameter, the resource cluster and the charging and discharging control instruction are visualized.

[0013] Preferably, the process of obtaining the preprocessed video stream data and the battery state data comprises:

[0014] A video acquisition device is deployed on each charging pile, and the position of the video acquisition device is opposite to the charging port area of the accessed electric vehicle, so as to obtain continuous video stream data of the charging port area;

[0015] A data acquisition unit is integrated in the control system of the charging pile, and the data acquisition unit establishes a communication connection with the battery management system of the accessed electric vehicle through a charging communication protocol, so as to obtain battery state data of the accessed electric vehicle, including battery state of charge data and battery surface temperature data;

[0016] Frame sampling processing is performed on the video stream data to obtain a video sequence;

[0017] Image transformation is performed on each frame of image in the video sequence to convert the multi-channel color image into a single-channel image;

[0018] The pixel values of the single-channel image are normalized to scale the pixel values to a standardized numerical interval to obtain preprocessed video stream data;

[0019] The collected battery state data is cleaned to process missing values and abnormal values;

[0020] A filtering algorithm based on a state estimation model is adopted to denoise the cleaned battery state data;

[0021] The denoised battery state data is standardized to eliminate the influence of the dimensions of different physical quantities, and preprocessed battery state data is obtained;

[0022] The timestamps of the preprocessed video stream data and the battery state data are aligned.

[0023] Preferably, the process of extracting the physical connection stability feature and establishing the early warning mechanism comprises:

[0024] The preprocessed video stream data is divided into video segments containing consecutive image frames according to a preset time step as input of the three-dimensional convolutional neural network model;

[0025] The three-dimensional convolutional neural network model extracts spatio-temporal features containing time and space dimensions by performing convolution operations on the video segments through three-dimensional convolution kernels;

[0026] The extracted spatio-temporal features are input into the fully connected regression layer of the three-dimensional convolutional neural network model, and a scalar value is output, which is defined as the physical connection stability feature;

[0027] A safety threshold of the physical connection stability is set, and the value of the physical connection stability feature is compared with the safety threshold in real time. When the value of the physical connection stability feature is lower than the safety threshold, an early warning signal is generated, and the corresponding electric vehicle is removed from the dispatchable resource pool.

[0028] Preferably, the process of extracting the thermal stress feature and the energy state feature from the preprocessed battery state data comprises:

[0029] The difference between the preprocessed battery surface temperature data at the current time and the preprocessed battery surface temperature data at the previous sampling time is calculated, and the difference is divided by the time interval between the two sampling times to obtain the thermal stress feature;

[0030] The energy state feature is extracted from the preprocessed battery state data, and the energy state feature is the battery state of charge data after standardization processing.

[0031] Preferably, the process of outputting the health state parameter of the battery comprises:

[0032] The thermal stress feature and the energy state feature are combined into a two-dimensional input feature vector at each sampling time;

[0033] The two-dimensional input feature vectors within a continuous time step are constructed into an input sequence with a time sequence relationship;

[0034] The input sequence is input into the gated recurrent unit network model, which processes the input sequence through internal update gates and reset gates, and updates the hidden state of the gated recurrent unit network model to capture the time dependence in the input sequence;

[0035] The hidden state output by the gated recurrent unit network model after processing the input sequence is input into the fully connected output layer;

[0036] The fully connected output layer outputs a two-dimensional vector, and the two-dimensional vector is defined as a health state parameter of the battery.

[0037] Preferably, the process of dividing the accessed electric vehicles into different resource clusters comprises:

[0038] The physical connection stability feature and the health state parameter of each accessed electric vehicle are combined into a three-dimensional data point, and the three-dimensional data points from each accessed electric vehicle collectively constitute a feature space;

[0039] A spatial clustering algorithm is applied to the density-based noise, and a neighborhood radius parameter and a core object minimum neighborhood point number parameter are set;

[0040] The Euclidean distance between two three-dimensional data points in the feature space is calculated;

[0041] For the three-dimensional data points in the feature space, the number of three-dimensional data points contained in the neighborhood of the three-dimensional data points within the neighborhood radius parameter range is determined by comparing the Euclidean distance with the neighborhood radius parameter;

[0042] If the number of three-dimensional data points contained in the neighborhood of the three-dimensional data points is not less than the core object minimum neighborhood point number parameter, the three-dimensional data points are marked as core objects;

[0043] Starting from an unvisited core object, a new cluster is created, density-reachable three-dimensional data points from the core object are searched, and the searched three-dimensional data points are divided into the new cluster;

[0044] The process of creating a new cluster from an unvisited core object, searching for density-reachable three-dimensional data points, and dividing the density-reachable three-dimensional data points into the new cluster is repeated until each core object is visited;

[0045] The set of electric vehicles corresponding to the three-dimensional data points in the cluster is defined as a resource cluster.

[0046] Preferably, the process of performing distributed collaborative solution to obtain the total power adjustment instruction for each resource cluster comprises:

[0047] The resource cluster is defined as an independent agent, and the agents constitute a multi-agent system;

[0048] A digital twin collaborative optimization model based on a multi-agent deep reinforcement learning algorithm is constructed;

[0049] An independent actor network is set for each agent, and a central critic network is set for the multi-agent system to form an actor-critic architecture;

[0050] The state space, action space, and reward function of the multi-agent system are defined;

[0051] The state space comprises aggregated parameters of each resource cluster and global information of the power grid, the aggregated parameters comprising average physical connection stability characteristics, average health state parameters, total charging power and total available capacity of the electric vehicles within the cluster;

[0052] The action space of each agent is a continuous power adjustment range of the resource cluster;

[0053] The reward function is used to evaluate the immediate reward obtained after each agent performs an action during the training process, and the calculation of the immediate reward is based on the power grid frequency deviation and the expected battery loss of the electric vehicles within the cluster;

[0054] In the digital twin collaborative optimization model, the parameters of the actor network and the critic network are updated through continuous interaction between the multi-agent system and the digital twin power grid environment until the digital twin collaborative optimization model converges;

[0055] The state space at the current time is input into the trained actor network, and the actor network of each agent selects and outputs an action from the action space, which is the total power adjustment instruction for the resource cluster.

[0056] Preferably, the process of performing the secondary power distribution and generating the charging and discharging control instructions comprises:

[0057] Within the resource cluster receiving the total power adjustment instruction, the normalized distribution weight coefficient is calculated according to the health state parameter and the physical connection stability characteristic of each electric vehicle;

[0058] The total power adjustment instruction is distributed to each electric vehicle in the resource cluster according to the normalized distribution weight coefficient, and the charging and discharging control instruction for each electric vehicle is generated;

[0059] The charging and discharging control instruction is issued to the corresponding charging pile execution end for execution;

[0060] The battery state data updated after the charging and discharging control instruction is executed is obtained in real time through the data acquisition unit, and the updated battery state data is used to update the health state parameter and the resource cluster to form the state space at the next time, forming a closed loop control.

[0061] Preferably, the process of visualizing the physical connection stability characteristics, health state parameters, resource clusters and charging and discharging control instructions comprises:

[0062] A digital twin three-dimensional topology view is constructed, which is used to display the geographical distribution and connection relationship of the charging infrastructure and the connected electric vehicles;

[0063] In the digital twin three-dimensional topology view, an interactive three-dimensional model is created for each connected electric vehicle.

[0064] Data binding is performed between the physical connection stability features, the health state parameters and the charging and discharging control instructions and corresponding three-dimensional models;

[0065] By integrating a time sequence curve diagram in the digital twin three-dimensional topological view, a change trend of the physical connection stability features over time is displayed;

[0066] By integrating a data dashboard in the digital twin three-dimensional topological view, real-time numerical values of the health state parameters and the charging and discharging control instructions are displayed;

[0067] In the digital twin three-dimensional topological view, a preset color mapping rule is adopted to color the three-dimensional models in real time according to the resource clusters to which each electric vehicle belongs, so as to visualize the results of dynamic clustering.

[0068] The virtual power plant vehicle network resource optimization control system based on digital twinning is used to implement the above method, and comprises:

[0069] A twin state acquisition module is configured to acquire and pre-process video stream data of a charging port area and battery state data of aggregated electric vehicles, to obtain pre-processed video stream data and battery state data;

[0070] A physical health feature extraction module is configured to input the pre-processed video stream data into a three-dimensional convolutional neural network model, extract physical connection stability features, and establish an early warning mechanism, and extract thermal stress features and energy state features from the pre-processed battery state data;

[0071] A battery health evaluation module is configured to input the thermal stress features and the energy state features into a gated recurrent unit network model, and output health state parameters of the battery;

[0072] A dynamic resource clustering module is configured to divide the accessed electric vehicles into different resource clusters by using a density-based noise application spatial clustering algorithm according to the physical connection stability features and the health state parameters;

[0073] A collaborative optimization decision module is configured to take aggregated parameters of the resource clusters as inputs of a digital twin collaborative optimization model based on a multi-agent deep reinforcement learning algorithm, perform distributed collaborative solving, obtain power adjustment total amount instructions for each resource cluster, and perform power secondary distribution to generate charging and discharging control instructions;

[0074] A state monitoring visualization module is configured to visualize the physical connection stability features, the health state parameters, the resource clusters and the charging and discharging control instructions.

[0075] The present application has the following advantages:

[0076] The application realizes real-time quantitative evaluation of the physical connection state of the charging interface by introducing analysis of the video stream data of the charging port area, establishes an early warning mechanism, avoids issuing scheduling instructions to electric vehicles with unreliable physical connections, and improves the success rate of scheduling instruction execution and the operational safety of the system.

[0077] The application realizes differentiated and refined control of batteries with different health states by accurately evaluating the battery health state of each electric vehicle and dynamically clustering resources according to the evaluation results, avoids damage to the battery caused by non-differential charging and discharging control, and effectively delays the process of battery capacity degradation.

[0078] The application realizes global optimization scheduling of large-scale heterogeneous vehicle network resources by constructing a digital twin collaborative optimization model that combines physical connection reliability and battery health state parameters and using a multi-agent deep reinforcement learning algorithm for distributed solution, and improves the control accuracy and response reliability of the virtual power plant as a whole. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 The flowchart of the method of the application;

[0080] Figure 2 The structural block diagram of the system of the application. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the application will be described clearly and completely below.

[0082] Embodiment 1: As shown in the figure, the virtual power plant vehicle network resource optimization control method based on digital twinning includes the following steps: Figure 1

[0083] S1, collect and pre-process the video stream data of the charging port area and the battery state data of the aggregated electric vehicles to obtain pre-processed video stream data and battery state data.

[0084] Deploy a video acquisition device on each charging pile, and the position of the video acquisition device is opposite to the charging port area of the connected electric vehicle, for acquiring continuous video stream data of the charging port area.

[0085] Integrate a data acquisition unit in the control system of the charging pile, and the data acquisition unit establishes a communication connection with the battery management system of the connected electric vehicle through a charging communication protocol, for acquiring battery state data of the connected electric vehicle, including battery state of charge data and battery surface temperature data.

[0086] Pre-process the video stream data:

[0087] Frame sampling processing is performed on the video stream data to obtain a video sequence;​

[0088] performing image transformation on each image in the video sequence to convert the multi-channel color image into a single-channel image;

[0089] performing normalization on the pixel value of the single-channel image to scale the pixel value into a standardized numerical interval, to obtain preprocessed video stream data.

[0090] performing preprocessing on the battery state data:

[0091] performing data cleaning on the collected battery state data to process missing values and abnormal values;

[0092] adopting a filtering algorithm based on a state estimation model to perform denoising processing on the cleaned battery state data;

[0093] performing standardization processing on the denoised battery state data to eliminate the influence of the dimensions of different physical quantities, to obtain preprocessed battery state data.

[0094] aligning the timestamps of the preprocessed video stream data and the battery state data.

[0095] S2, inputting the preprocessed video stream data into a three-dimensional convolutional neural network model, extracting physical connection stability features, and establishing a warning mechanism to extract thermal stress features and energy state features from the preprocessed battery state data.

[0096] segmenting the preprocessed video stream data into video segments containing consecutive image frames according to a preset time step, as input of the three-dimensional convolutional neural network model;

[0097] the three-dimensional convolutional neural network model performs convolution operation on the video segments through a three-dimensional convolution kernel to extract spatio-temporal features containing time dimension and space dimension;

[0098] inputting the extracted spatio-temporal features into a fully connected regression layer of the three-dimensional convolutional neural network model to output a scalar value, and defining the scalar value as a physical connection stability feature;

[0099] setting a safety threshold of the physical connection stability, and comparing the value of the physical connection stability feature with the safety threshold in real time, when the value of the physical connection stability feature is lower than the safety threshold, generating a warning signal, and removing the corresponding electric vehicle from the dispatchable resource pool;

[0100] calculating the difference between the preprocessed battery surface temperature data at the current time and the preprocessed battery surface temperature data at the previous sampling time, and dividing the difference by the time interval between the two sampling times to obtain the thermal stress feature;

[0101] The energy state feature is extracted from the pretreated battery state data, and the energy state feature is the battery state of charge data after standardization processing.

[0102] S3, inputting the thermal stress feature and the energy state feature into a gated recurrent unit network model to output a health state parameter of the battery.

[0103] The thermal stress feature and the energy state feature are combined into a two-dimensional input feature vector at each sampling time;

[0104] The two-dimensional input feature vectors in the continuous time steps are constructed into an input sequence with a time sequence relationship;

[0105] The input sequence is input into the gated recurrent unit network model, and the gated recurrent unit network model processes the input sequence through the internal update gate and reset gate, and updates the hidden state of the gated recurrent unit network model to capture the time dependence in the input sequence;

[0106] The reset gate in the gated recurrent unit network model determines the degree of retaining the hidden state information of the previous time to the current time;

[0107] The update gate in the gated recurrent unit network model determines the inclusion of the hidden state information of the previous time and the new input information of the current time in the hidden state of the current time;

[0108] The hidden state output by the gated recurrent unit network model after processing the input sequence is input into a fully connected output layer;

[0109] The fully connected output layer outputs a two-dimensional vector, the first dimension value of the two-dimensional vector represents the current capacity attenuation degree of the battery, and the second dimension value represents the current internal resistance increase degree of the battery, and the two-dimensional vector is defined as the health state parameter of the battery.

[0110] S4, according to the physical connection stability feature and the health state parameter, adopting a density-based noise application spatial clustering algorithm, the accessed electric vehicles are divided into different resource clusters.

[0111] The physical connection stability feature and the health state parameter of each accessed electric vehicle are combined into a three-dimensional data point, and the three-dimensional data points from each accessed electric vehicle together constitute a feature space;

[0112] For the density-based noise application spatial clustering algorithm, the neighborhood radius parameter and the core object minimum neighborhood point number parameter are set;

[0113] The Euclidean distance between two three-dimensional data points in the feature space is calculated;

[0114] For a three-dimensional data point in the feature space, by comparing the Euclidean distance with the neighborhood radius parameter, it is determined that the number of three-dimensional data points contained in the neighborhood of the three-dimensional data point within the neighborhood radius parameter range;

[0115] If the number of three-dimensional data points contained in the neighborhood of the three-dimensional data point is not less than the core object minimum neighborhood point number parameter, the three-dimensional data point is marked as a core object;

[0116] Starting from an unvisited core object, a new cluster is created, density-reachable three-dimensional data points from the core object are searched for, and the searched three-dimensional data points are divided into the new cluster;

[0117] The process of starting from an unvisited core object to create a new cluster, searching for density-reachable three-dimensional data points, and dividing the density-reachable three-dimensional data points into the new cluster is repeated until the access to each core object is completed;

[0118] The set of electric vehicles corresponding to the three-dimensional data points in the cluster is defined as a resource cluster.

[0119] S5, taking the aggregation parameters of the resource cluster as the input of the digital twin collaborative optimization model based on the multi-agent deep reinforcement learning algorithm, performing distributed collaborative solution, obtaining the total power adjustment instruction for each resource cluster, and performing power secondary allocation to generate the charging and discharging control instruction.

[0120] The resource cluster is defined as an independent agent, and the agents constitute a multi-agent system;

[0121] A digital twin collaborative optimization model based on a multi-agent deep reinforcement learning algorithm is constructed;

[0122] An independent actor network is set for each agent, and a central critic network is set for the multi-agent system to form an actor-critic architecture;

[0123] The state space, action space and reward function of the multi-agent system are defined;

[0124] The state space includes the aggregation parameters of each resource cluster and the global information of the power grid, and the aggregation parameters include the average physical connection stability feature, the average health state parameter, the total charging power and the total available capacity of the electric vehicles in the cluster;

[0125] The action space of each agent is the continuous power adjustment range of the resource cluster;

[0126] The reward function is used to evaluate the immediate reward obtained after each agent performs an action during the training process, and the calculation basis of the immediate reward is the power frequency deviation and the expected battery loss of the electric vehicles in the cluster;

[0127] In the digital twin collaborative optimization model, through continuous interaction of the multi-agent system and the digital twin power grid environment, the parameters of the actor network and the critic network are updated until the digital twin collaborative optimization model converges.

[0128] The state space at the current time is input into the trained actor network, and the actor network of each agent selects and outputs an action from the action space, which is the total power adjustment instruction for the resource cluster.

[0129] In the resource cluster receiving the total power adjustment instruction, the normalized allocation weight coefficient is calculated according to the health state parameter and the physical connection stability feature of each electric vehicle.

[0130] The total power adjustment instruction is allocated to each electric vehicle in the resource cluster according to the normalized allocation weight coefficient, and the charge-discharge control instruction for each electric vehicle is generated.

[0131] The charge-discharge control instruction is executed by the corresponding charging pile execution end.

[0132] Through the data acquisition unit, the updated battery state data after executing the charge-discharge control instruction is obtained in real time, and the updated battery state data is used to update the health state parameter and the resource cluster to form the state space at the next time, forming a closed loop control.

[0133] S6, visualize the physical connection stability feature, the health state parameter, the resource cluster and the charge-discharge control instruction.

[0134] A digital twin three-dimensional topology view is constructed, which is used to show the geographical distribution and connection relationship of charging infrastructure and connected electric vehicles.

[0135] In the digital twin three-dimensional topology view, an interactive three-dimensional model is created for each connected electric vehicle.

[0136] The physical connection stability feature, the health state parameter and the charge-discharge control instruction are data bound with the corresponding three-dimensional model.

[0137] By integrating the time series chart in the digital twin three-dimensional topology view, the trend of the physical connection stability feature over time is shown.

[0138] By integrating the data dashboard in the digital twin three-dimensional topology view, the real-time values of the health state parameter and the charge-discharge control instruction are shown.

[0139] In the digital twin three-dimensional topology view, a preset color mapping rule is used to color the three-dimensional model in real time according to the resource cluster to which each electric vehicle belongs, to visualize the dynamic clustering result.

[0140] Embodiment 2: As shown in Figure 2 The virtual power plant vehicle network resource optimization control system based on digital twinning is used to implement the method in embodiment 1, and comprises:

[0141] A twin state acquisition module is configured to acquire and pre-process video stream data of a charging port area and battery state data of aggregated electric vehicles to obtain pre-processed video stream data and battery state data.

[0142] A physical health feature extraction module is configured to input the pre-processed video stream data into a three-dimensional convolutional neural network model, extract physical connection stability features, and establish an early warning mechanism, and extract thermal stress features and energy state features from the pre-processed battery state data.

[0143] A battery health evaluation module is configured to input the thermal stress features and energy state features into a gated recurrent unit network model to output health state parameters of the battery.

[0144] A dynamic resource clustering module is configured to divide the accessed electric vehicles into different resource clusters by using a density-based noise application spatial clustering algorithm according to the physical connection stability features and the health state parameters.

[0145] A collaborative optimization decision module is configured to take the aggregated parameters of the resource clusters as inputs of a digital twinning collaborative optimization model based on a multi-agent deep reinforcement learning algorithm, perform distributed collaborative solving, obtain power adjustment total amount instructions for each resource cluster, and perform power secondary distribution to generate charging and discharging control instructions.

[0146] A state monitoring visualization module is configured to visualize the physical connection stability features, the health state parameters, the resource clusters, and the charging and discharging control instructions.

Claims

1. A virtual power plant vehicle network resource optimization control method based on digital twinning, characterized in that, The method comprises the following steps: Collect and pre-process the video stream data of the charging port area and the battery state data of the electric vehicle to obtain pre-processed video stream data and battery state data, and the battery state data includes battery state of charge data and battery surface temperature data; Input the pre-processed video stream data into a three-dimensional convolutional neural network model, extract the physical connection stability feature, and establish a warning mechanism, the process comprising: Divide the pre-processed video stream data into video segments containing consecutive image frames according to the preset time step as the input of the three-dimensional convolutional neural network model; The three-dimensional convolutional neural network model extracts spatio-temporal features containing time and space dimensions through three-dimensional convolution kernel convolution operation on the video segment; Input the extracted spatio-temporal features into the fully connected regression layer of the three-dimensional convolutional neural network model to output a scalar value, and define the scalar value as the physical connection stability feature; Set a safety threshold for the physical connection stability, and compare the value of the physical connection stability feature with the safety threshold in real time, and when the value of the physical connection stability feature is lower than the safety threshold, generate a warning signal, and remove the corresponding electric vehicle from the schedulable resource pool; Extract the thermal stress feature and the energy state feature from the pre-processed battery state data, the process comprising: Calculate the difference between the pre-processed battery surface temperature data at the current time and the pre-processed battery surface temperature data at the previous sampling time, and divide the difference by the time interval between the two sampling times to obtain the thermal stress feature; Extract the energy state feature from the pre-processed battery state data, and the energy state feature is the battery state of charge data after standardization processing; Input the thermal stress feature and the energy state feature into the gated recurrent unit network model to output the health state parameter of the battery; According to the physical connection stability feature and the health state parameter, a spatial clustering algorithm based on density-based noise application is used to divide the accessed electric vehicles into different resource clusters; Take the aggregated parameters of the resource clusters as the input of the digital twin collaborative optimization model based on the multi-agent deep reinforcement learning algorithm, and perform distributed collaborative solution to obtain the total power adjustment instruction for each resource cluster, and perform power secondary distribution to generate the charging and discharging control instruction; Visualize the physical connection stability feature, the health state parameter, the resource cluster, and the charging and discharging control instruction.

2. The digital-twin-based virtual power plant vehicle-to-grid resource optimization control method according to claim 1, characterized in that, The process of obtaining the pre-processed video stream data and the battery state data comprises: Deploy a video acquisition device on each charging pile, and the position of the video acquisition device is opposite to the charging port area of the accessed electric vehicle, which is used to obtain continuous video stream data of the charging port area; Integrate a data acquisition unit in the control system of the charging pile, and the data acquisition unit establishes a communication connection with the battery management system of the accessed electric vehicle through a charging communication protocol to obtain the battery state data of the accessed electric vehicle, including the battery state of charge data and the battery surface temperature data; Frame sampling processing is performed on the video stream data to obtain a video sequence; Image transformation is performed on each frame of image in the video sequence to convert the multi-channel color image into a single-channel image; The pixel values of the single-channel image are normalized, and the pixel values are scaled to a standardized numerical interval to obtain pre-processed video stream data; The collected battery state data is cleaned to handle missing values and outliers; A filtering algorithm based on a state estimation model is used to denoise the cleaned battery state data; The denoised battery state data is standardized to eliminate the dimension influence of different physical quantities, and pre-processed battery state data is obtained; The timestamps of the pre-processed video stream data and the battery state data are aligned.

3. The digital-twin-based virtual power plant vehicle-to-grid resource optimization control method according to claim 1, characterized in that, The process of outputting the health state parameter of the battery includes: combining the thermal stress feature and the energy state feature into a two-dimensional input feature vector at each sampling time; constructing the two-dimensional input feature vector within a continuous time step into an input sequence with a time sequence relationship; inputting the input sequence into the gated recurrent unit network model, which processes the input sequence through the internal update gate and reset gate, and updates the hidden state of the gated recurrent unit network model to capture the time dependence in the input sequence; inputting the hidden state output by the gated recurrent unit network model after processing the input sequence into the fully connected output layer; the fully connected output layer outputs a two-dimensional vector, and the two-dimensional vector is defined as the health state parameter of the battery.

4. The digital-twin-based virtual power plant vehicle-to-grid resource optimization control method according to claim 1, characterized in that, The process of dividing the accessed electric vehicles into different resource clusters includes: combining the physical connection stability feature and the health state parameter of each accessed electric vehicle into a three-dimensional data point, and collectively constructing the three-dimensional data points from each accessed electric vehicle into a feature space; applying a density-based noise spatial clustering algorithm, setting a neighborhood radius parameter and a core object minimum neighborhood point number parameter; calculating the Euclidean distance between two three-dimensional data points in the feature space; for the three-dimensional data points in the feature space, by comparing the Euclidean distance with the neighborhood radius parameter, determining the number of three-dimensional data points contained in the neighborhood of the three-dimensional data points within the neighborhood radius parameter range; if the number of three-dimensional data points contained in the neighborhood of the three-dimensional data point is not less than the core object minimum neighborhood point number parameter, the three-dimensional data point is marked as a core object; starting from an unvisited core object, creating a new cluster, finding three-dimensional data points density reachable from the core object, and dividing the found three-dimensional data points into the new cluster; repeating the process of starting from an unvisited core object to create a new cluster, finding density reachable three-dimensional data points, and dividing the density reachable three-dimensional data points into the new cluster until each core object is accessed; the set of electric vehicles corresponding to the three-dimensional data points in the cluster is defined as a resource cluster.

5. The digital-twin-based virtual power plant vehicle-to-grid resource optimization control method according to claim 1, characterized in that, The process of distributed collaborative solving to obtain the total power adjustment instruction for each resource cluster includes: defining the resource cluster as an independent agent, and the agents constitute a multi-agent system; building a digital twin collaborative optimization model based on a multi-agent deep reinforcement learning algorithm; setting an independent actor network for each agent, and setting a central critic network for the multi-agent system to form an actor-critic architecture; defining the state space, action space and reward function of the multi-agent system; The state space comprises aggregated parameters of each resource cluster and global information of the power grid, and the aggregated parameters comprise average physical connection stability characteristics, average health state parameters, total charging power and total available capacity of the electric vehicles in the cluster; The action space of each agent is a continuous power adjustment range of the resource cluster; The reward function is used to evaluate an immediate reward obtained after each agent performs an action in the training process, and the immediate reward is calculated based on a power grid frequency deviation and an expected battery loss of the electric vehicles in the cluster; In the digital twin collaborative optimization model, the multi-agent system is trained through continuous interaction with the digital twin power grid environment, and parameters of the actor network and the critic network are updated until the digital twin collaborative optimization model converges; The state space at the current time is input into the trained actor network, and each agent selects and outputs an action from the action space through the actor network, and the action is a total power adjustment instruction for the resource cluster.

6. The digital-twin-based virtual power plant vehicle-to-grid resource optimization control method according to claim 1, characterized in that, The process of performing the secondary power distribution to generate the charging and discharging control instruction comprises: Inside the resource cluster receiving the total power adjustment instruction, normalized distribution weight coefficients are calculated according to the health state parameters and the physical connection stability characteristics of each electric vehicle; The total power adjustment instruction is distributed to each electric vehicle in the resource cluster according to the normalized distribution weight coefficients to generate a charging and discharging control instruction for each electric vehicle; The charging and discharging control instruction is sent to the corresponding charging pile execution end for execution; The battery state data updated after the charging and discharging control instruction is executed is acquired in real time through the data acquisition unit, and the updated battery state data is used to update the health state parameters and the resource cluster to form a state space at the next time, forming a closed loop control.

7. The digital-twin-based virtual power plant vehicle-to-grid resource optimization control method according to claim 1, characterized in that, The process of visualizing the physical connection stability characteristics, the health state parameters, the resource cluster and the charging and discharging control instruction comprises: A digital twin three-dimensional topology view is constructed, and the digital twin three-dimensional topology view is used to display the geographical distribution and connection relationship of the charging infrastructure and the connected electric vehicles; An interactive three-dimensional model is created for each connected electric vehicle in the digital twin three-dimensional topology view; The physical connection stability characteristics, the health state parameters and the charging and discharging control instruction are data-bound to the corresponding three-dimensional model; The trend of the physical connection stability characteristics over time is displayed through a time series chart integrated in the digital twin three-dimensional topology view; Real-time values of the health state parameters and the charging and discharging control instruction are displayed through a data dashboard integrated in the digital twin three-dimensional topology view; In the digital twin three-dimensional topology view, the three-dimensional models are colored in real time according to the resource clusters to which each electric vehicle belongs, so as to visualize the results of dynamic clustering.

8. A virtual power plant vehicle grid resource optimization control system based on digital twinning, for implementing the method of any one of claims 1-7, characterized in that, The process of visualizing the physical connection stability characteristics, the health state parameters, the resource cluster and the charging and discharging control instruction comprises: A digital twin state acquisition module is configured to acquire and pre-process video stream data of a charging port area and battery state data of aggregated electric vehicles to obtain pre-processed video stream data and battery state data. The physical health feature extraction module is configured to input the preprocessed video stream data into a three-dimensional convolutional neural network model, extract a physical connection stability feature, and establish a pre-warning mechanism to extract a thermal stress feature and an energy state feature from the preprocessed battery state data; The battery health evaluation module is configured to input the thermal stress feature and the energy state feature into a gated recurrent unit network model, and output a health state parameter of the battery; The dynamic resource clustering module is configured to divide the accessed electric vehicles into different resource clusters by using a density-based noise application spatial clustering algorithm according to the physical connection stability feature and the health state parameter; The collaborative optimization decision module is configured to take the aggregated parameters of the resource clusters as inputs of a digital twin collaborative optimization model based on a multi-agent deep reinforcement learning algorithm, perform distributed collaborative solving, obtain a total power adjustment instruction for each resource cluster, and perform secondary power distribution to generate a charging and discharging control instruction; The state monitoring visualization module is configured to visualize the physical connection stability feature, the health state parameter, the resource cluster and the charging and discharging control instruction.

Images